OpenFold3 debuts as Apache-licensed protein co-folding model, with Apheris bringing it into pharma IT

The OpenFold Consortium released a preview of OpenFold3, an Apache 2.0-licensed foundation model that predicts 3D protein structures and co-folded complexes with small-molecule ligands, nucleic acids and other proteins, positioning it for industry use at scale.

Simultaneously, Apheris, a Berlin-based provider of enterprise-grade AI applications for drug discovery, launched ApherisFold to run, benchmark, and fine-tune OpenFold3 securely inside pharma environments, including on-prem and private cloud. The software supports secure local inference, benchmarking on public or proprietary data, federated fine-tuning, and GUI/API access, with deployment via AWS CloudFormation or pre-built containers.

“With ApherisFold, we’re turning open science into a production-ready product that can be used by enterprises—giving pharma scientists instant, secure access to the world’s most powerful co-folding models,” said Robin Röhm, Apheris co-founder and CEO.

ApherisFold’s GUI gives researchers detailed structural insights through interactive 3D models and confidence scoring:

Unlike AlphaFold3’s limited non-industrial license, OpenFold3’s Apache 2.0 license permits companies to test, train, adapt and build applications on the model, opening the door to broad commercial adoption.

Apheris leads a federated OpenFold3 initiative through the AI Structural Biology (AISB) Network with five of the top-20 pharmas—AbbVie, Astex Pharmaceuticals, Bristol Myers Squibb, Johnson & Johnson, and Takeda. “OpenFold3 is now on par with AlphaFold3 for protein–ligand co-folding and as such significantly outperforming Boltz-2,” Röhm explained in a message to R&D World. The federated approach trains the model on roughly five times more proprietary structural data directly relevant to industrial drug discovery compared to the Protein Data Bank.

OpenFold3 was trained on more than 300,000 experimentally determined structures plus an OpenFold-curated synthetic set of more than 40 million. Built with PyTorch and packaged as a NVIDIA NIM microservice for scalable inference, the model’s modular design integrates into existing pipelines.

Early adopters include Novo Nordisk, Bayer Crop Science, Outpace Bio, and Cyrus Biotechnology, spanning use cases from small-molecule discovery to engineered cell-therapy circuits. The UK’s OpenBind initiative will fine-tune OpenFold3 using collaborator-generated data. Access options include preview code on GitHub, a Docker image and checkpoint on Hugging Face, hosted access via Tamarind Bio, and locally deployable versions through Apheris.